Knowledge graph assisted extreme summarization of scientific documents in hyperbolic space

Abstract

A research paper is a document that presents an original work and introduces new concepts and makes interconnections between them via arguments and statements. Extreme Summarization involves high compression of the information content of the document and representing it in a concise, meaningful format. Hence extreme summarization of scientific documents entails representation of the key concepts as presented in the concerned document in a concise and coherent manner.The challenge of this task is to capture the essential concepts as presented in the entire paper. In this paper we handle the problem of abstractive extreme summarization of scientific documents.The technique of abstractive summarization would allow us to concisely represent the information present in the entire scientific research paper by generating a new sentence rather than being constrained to having to pick sentences that is already present in the document. Since no single sentence in the document can capture the entire information presented in the document. In order to do effective extreme summarization of scientific documents,we propose Knowledge graph assisted hyperbolic BART(KAHB) , a knowledge graph assisted sequence to sequence architecture while transforming the intermediate embeddings into hyperbolic space. A knowledge graph helps to capture the interconnection between the concepts as presented in the paper, which a plain sequence to sequence model fails to do. Transforming the embeddings to an hyperbolic space helps to capture the inherent hierarchical relationships present in the document .Applying the above methods , we are able to achieve improvements in the performance of a standard sequence to sequence mode.